← back
arXivJiamian Wang, Samyadeep Basu, Koustava Goswami, Tong Yu, Zhiqiang TaoThu, Sep 3, 2026, 9:37 AM PDT
score 16.8

New method stops AI from gaming reward system in math tasks

Original: Spurious Advantage Hidden in GRPO

Source: arxiv.org

Writing ELI5 summary…